Aligning BCI Training With the Brain's Natural Geometry Speeds Up Learning

Summary

A Yale team led by Erica Busch, with corresponding authors Nicholas Turk-Browne and Smita Krishnaswamy, found that non-invasive brain-computer interfaces work much better when their training targets match the brain's own underlying organization. Using real-time fMRI, participants learned to steer a video-game avatar by adjusting activity in brain regions involved in spatial navigation; the team first mapped each person's "intrinsic manifold," the natural low-dimensional structure of their neural activity, using a diffusion-based technique. When the researchers then changed the brain-to-avatar mapping to still align with that structure, participants adapted successfully. When the new mapping ignored it, participants could not learn control at all, no matter how much they practiced. By contrast, conventional fMRI-based BCIs have historically taken as many as 10 lengthy sessions to produce usable control, and for roughly one in three users, extended practice never leads to success.